When Explaining Is Governing: Comprehensibility for the AI Era

The views expressed are those of the author and do not necessarily reflect the views of ASPA as an organization.

By Mauricio Covarrubias
October 10, 2025

In the age of artificial intelligence, governments do not earn legitimacy by posting data alone but by offering comprehensible reasons for decisions that affect people’s lives. Transparency made the state visible; comprehensibility makes it answerable. It turns data and rules into public reason that residents can understand, contest and when appropriate, correct.

Administrative decisions increasingly rely on algorithmic systems—risk scores, prioritization engines, eligibility models. In this context, explanation is not a communications add-on; it is governing with due process, making criteria, limits and real avenues for review explicit. Visibility without understanding produces spectators; reason-giving with action paths produces citizens.

Second-Order Opacity

The open-data decade solved access while leaving a subtler opacity intact: datasets are open yet decision rules—thresholds, weights, business logic—remain implicit. As Zachary C. Lipton argued in The Mythos of Model Interpretability, technical talk often stalls at averages and abstractions that say little about individual cases. Meanwhile, the survey of explainable AI by Arrieta and colleagues shows that opening models is insufficient if explanations are not useful to those impacted. People want to reconstruct the why in their own case and learn what they can do next.

Comprehensibility: When Explaining Is Governing

To explain is not to paraphrase jargon; it is to enable reconstruction of the reasoning that connects data, rules and outcomes—an ambition that Finale Doshi-Velez and Been Kim called for in a more rigorous science of interpretability. In civic settings, such reconstruction works best across three coordinated layers: a citizen layer (what was decided about me, why and what I can do—correct, provide evidence, appeal), a technical layer (influential variables, performance and fairness tests) and a governance layer (accountable officials, audits and version control for rules and models). The goal is not to “technify” the public but to return agency to affected individuals.

Explicability ≠ Comprehensibility (And Why It Matters)

The terms are often conflated. Explicability faces the system—how the model works, which variables matter and what data plus performance/fairness metrics support it—anchored by Model Cards (Mitchell et al.) and Datasheets for Datasets (Gebru et al.). Comprehensibility faces the person and the case: a situated narrative (“what happened and why in my file”), a clear counterfactual (Wachter, Mittelstadt, Russell) showing the minimal change to alter the outcome, and concrete action routes with deadlines and responsible offices. They are complementary: without explicability there’s no verifiable base; without comprehensibility there’s no actionable due process. The first illuminates computation; the second turns reasons into rights.

Interfaces are political: form is substance. A portal that informs without enabling action manufactures spectators; an interface that offers reasons and routes to review builds citizenship. In practice, the trio Model Cards + Datasheets + Counterfactuals functions as a working grammar for agencies that treat explaining as governing.

Crucially, reason-giving is not naïve. Some fear that explanations invite gaming. The concern is legitimate, but it does not cancel the democratic duty to justify public decisions. From a public-law perspective, Mark Bovens reminds us that accountability requires forums, standards and the possibility of sanction; burying criteria under opacity weakens that architecture. On the legal-technical front, Joshua Kroll and co-authors show that traceability—knowing which rule or model version applied—is a condition for auditing algorithms. And Andrew Selbst with Solon Barocas warn against the seductive but unrealistic idea of “explaining everything,” arguing instead for proportionality: the higher a decision’s stakes, the stronger the duty to give reasons and the broader the avenues for review.

From Portal to “Why”: A Simple Case

Scholarships illustrate the shift. Transparency posts a PDF of IDs, scores and a criteria annex. Comprehensibility lets each applicant look up their case and see a concise narrative: the applied rule, key variables, the threshold, a nearby counterfactual (e.g., “with a 3.4 instead of 3.2 you clear the cutoff”) and concrete action routes (correct records, upload documents, request review) with deadlines and responsible offices. The page links to the Model Card and a public changelog of criteria/thresholds, and the agency publishes quarterly performance and fairness metrics with comparators (history, peers, legal standards). This format works whether decisions follow traditional rules or algorithmic models.

Measure what matters

If we keep rewarding portals for pageviews and downloads, we will optimize visibility—not understanding. Managers should track indicators aligned with due process: perceived clarity (a short post-interaction survey), task success (completing a correction or appeal unaided), time-to-action, informed appeals rate (an increase can signal a more capable public) and version traceability (reconstructing the exact rule/model applied to a case). These are not vanity metrics; they are institutional learning levers.

Closing: A Realistic Itinerary

No shortcuts—just a clear path: inventory decisions that classify/allocate; risk-prioritize high-impact ones; prototype three minimum artifacts with real users (reason-giving notice, Model Card, rules/models changelog); institutionalize the changelog; and measure what matters (clarity, task success, time-to-action, informed appeals). The payoff is concrete: less confusion, earlier corrections, stronger organizational learning and above all, trust. In the AI era, explaining is governing: public reasons must be understandable, debatable and auditable, in the spirit of Sabel & Zeitlin’s experimentalist governance.

---

Author: Mauricio Covarrubias is Professor at the National Institute of Public Administration in Mexico. He holds a Ph.D. from the National Autonomous University of Mexico (UNAM) and completed a postdoctoral fellowship in Government and Public Policy at the University of New Mexico in Albuquerque. He is co-founder of the International Academy of Political-Administrative Sciences (IAPAS). He can be reached at [email protected] and followed on X (formerly Twitter) @OMCovarrubias.

(6 votes, average: 5.00 out of 5)
Loading...